It's been discussed and argued, and stated by Analog Devices's experts with the LTspice code, that LTspice should be entirely deterministic, with no randomness in simulations unless the user actually adds it.? One of the tools sometimes used to solve systems of equations, is adding a little randomness into the mix.? The randomness can steer the solver one way or the other, thus potentially avoiding difficult solutions.? But we are assured that LTspice does not do that.? As far as I am aware, SPICE itself (from Berkeley) did not either.? As far as I know, "numerical noise" also does not exist in the sense that repeated runs through the same code should be exactly the same, down to the very last bit.? Even "round-off errors" are exactly predictable and repeatable.
?
If I remember correctly, ADI also said that LTspice code did have some unintended non-determinism, and that these have been (or are being) cleaned up in recent LTspice releases.
?
Over the years I have seen cases where LTspice simulated differently on repeated runs.? We know it should not have done that.? My guess is that I was witnessing some of those unintended bugs that ADI has been fixing.
?
Maybe they are not all fixed yet.? (And as they say, you can't get rid of ALL bugs.)
?
Or maybe you (Jerry Lee Marcel) had certain settings set, the first time you tried this simulation, which caused your simulation to converge correctly right "out of the box" the first time, but not a day later when you fired up LTspice again and it had reverted back to your default settings.? Who knows?
?
Andy
?
